Reproduce “Zero-Shot Logit Adjustment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dubing Chen, Yuming Shen, Haofeng Zhang, Philip H.S. Torr
The authors propose ZLAP for generalized zero-shot recognition, where classifiers predict seen and unseen classes. Rather than rely on generated unseen features, ZLAP uses a prior to adjust classifier decision boundaries, while semantic prototypes carry class information. On AWA2, WGAN+ZLAP reaches 72.8% harmonic-mean accuracy versus 63.7% for f-CLSWGAN, using 10 generated features per unseen class; gains are less pronounced on SUN.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.